Adaptive smart card-based pull control systems in context-aware manufacturing systems: Training a neural network through multi-objective simulation optimization

Abstract

Nowadays, context aware manufacturing systems offer interesting capabilities to improve the performance of pull controlled production systems. Smart Kanbans can be used instead of physical cards, and the information become available about the production context, collected for example through sensors and RFID. Such information can be exploited by intelligent pull control strategies so as to dynamically adapt the number of cards. This is particularly useful for production systems that are subjected to unpredictable changes in the customers’ demand, and need to react quickly to preserve a high level of performance. For this reason, we aim, in this article, at proposing an intelligent system, which can communicate with the information system, whose purpose is to autonomously decide or to help managers in adding or removing cards. In this respect, we propose an approach that uses a neural network which is trained offline, directly from simulation, to decide when it is relevant to change the number of cards, and at what production stage. The learning process, based on multi-objective simulation optimization, aims at reducing the production costs as well as the number of changes to avoid nervousness. The use of stochastic simulation, allows various types of complex problems, related to manufacturing systems, to be addressed and fluctuating demand phenomena to be taken into account. The relevance of our approach is illustrated using six published adaptive ConWIP and Kanban systems. Comparisons with adaptive Kanban and ConWIP systems show that the neural network can automatically learn very relevant knowledge. Good results are obtained in terms of performance, with fewer changes in the number of cards. Several possible future research directions are pointed out.